On Thu, Jun 21, 2018 at 08:56:23PM +0300, Alexey Lyashkov wrote: > Majd, > > how soon? I have plan to work patch on next week, > i looks copy-paste of pointed Leon with some minor changes and applied to the different buffer. > Changing CQ was low hanging fruit because CQs are in multiple of CQEs (64bytes), for QPs you will need to take into account different sizes while allocating fragmented buffers. Thanks > > 21 июня 2018 г., в 14:20, Majd Dibbiny <majd@xxxxxxxxxxxx> написал(а): > > > > > >> On Jun 21, 2018, at 1:14 PM, Leon Romanovsky <leon@xxxxxxxxxx> wrote: > >> > >>> On Wed, Jun 20, 2018 at 10:48:16AM +0300, Alexey Lyashkov wrote: > >>> Hi All, > >>> > >>> while a lustre testing I have seen an very high ordered allocations was done by mlx5 driver. > >>> Similar bug reported agaist mlx4 driver in Lustre ticket https://jira.whamcloud.com/browse/LU-10736. > >>> As i see both fails related to the SGE array buffer allocation (sorry, i don’t know a good terms for it). > >>> But it uses a continues space, instead of fragmented. > >>> I have several questions about it > >>> What a reason for it? why don’t use a fragmented allocation and merge it logically as before for mlx4? > >> > >> I don't know the reasons for it, but we are working to avoid such large > >> allocations, for example see commit 88ca8be0037 "IB/mlx5: Implement fragmented > >> completion queue (CQ)" > > Soon we are going to do it for QPs of mlx5 and mlx4 as well. > >> > >> Thanks >
Attachment:
signature.asc
Description: PGP signature
